2024 VIS Area Curation Committee Executive Summary

Quantitative Analysis

Code
import itertools

import pandas as pd
import numpy as np

# Import the necessaries libraries
import plotly.offline as pio
import plotly.graph_objs as go
import plotly.express as px
# [jdf] no need to specify the renderer but, for interactive use, init_notebook should be called
# pio.renderers.default = "jupyterlab"
# Set notebook mode to work in offline
# pio.init_notebook_mode()
# pio.init_notebook_mode(connected=True)
width = 750

import sqlite3

#### Data Preparation

# static data – codes -> names etc.
staticdata = dict(
    decision = { 
        'C': 'Confer vs. cond Accept', # relevant for the 2020 and 2021 data have a different meaning
        'A': 'Accept', # for the 2020 data
        'A2': 'Accept', # after the second round, should be 120 in 2022
        'R': 'Reject', # reject after the first round -- should be 322 in 2022
        'R2': 'Reject in round 2', # reject after the second round -- should be 2 in 2022
        'R-2nd': 'Reject in round 2', 
        'DR-S': 'Desk Reject (Scope)', # should be 7 in 2022
        'DR-P': 'Desk Reject (Plagiarism)', # should be 4 in 2022
        'AR-P': 'Admin Reject (Plagiarism)', # should be 1 in 2022
        'DR-F': 'Desk Reject (Format)', # should be 4 in 2022
        'R-Strong': 'Reject Strong', # cannot resubmit to TVCG for a year
        'T': 'Reject TVCG fasttrack', # Explicitly invited to resubmit to TVCG, status in major revision
    },
    FinalDecision = { # Just flatten to Accept and Reject
        'C': 'Accept', 
        'A': 'Accept', # for the 2020 data
        'A2': 'Accept', # after the second round, should be 120 in 2022
        'R': 'Reject', # reject after the first round -- should be 322 in 2022
        'R2': 'Reject', # reject after the second round -- should be 2 in 2022
        'R-2nd': 'Reject', 
        'DR-S': 'Reject', # should be 7 in 2022
        'DR-P': 'Reject', # should be 4 in 2022
        'AR-P': 'Reject', # should be 1 in 2022
        'DR-F': 'Reject', # should be 4 in 2022
        'R-Strong': 'Reject',
        'T': 'Reject',
    },
    area = {
        'T&E': 'Theoretical & Empirical',
        'App': 'Applications',
        'S&R': 'Systems & Rendering',
        'R&I': 'Representations & Interaction',
        'DTr': 'Data Transformations',
        'A&D': 'Analytics & Decisions',
    },
    bid = { 
        0: 'no bid',
        1: 'want',
        2: 'willing',
        3: 'reluctant',
        4: 'conflict'
    },
    stat = {
        'Prim': 'Primary', 
        'Seco': 'Secondary'
    },
    keywords = pd.read_csv("../data/2021/keywords.csv", sep=';'), # 2021 is correct as there was no new keywords file in 2022
    colnames = {
        'confsubid': 'Paper ID',
        'rid': 'Reviewer',
        'decision': 'Decision',
        'area': 'Area',
        'stat': 'Role',
        'bid': 'Bid'
    }
)

dbcon = sqlite3.connect('../data/vis-area-chair.db') #[jdf] assume data is in ..

submissions_raw20 = pd.read_sql_query('SELECT * from submissions WHERE year = 2020', dbcon, 'sid')
submissions_raw21 = pd.read_sql_query('SELECT * from submissions WHERE year = 2021', dbcon, 'sid')
submissions_raw22 = pd.read_sql_query('SELECT * from submissions WHERE year = 2022', dbcon, 'sid')
submissions_raw23 = pd.read_sql_query('SELECT * from submissions WHERE year = 2023', dbcon, 'sid')
submissions_raw24 = pd.read_sql_query('SELECT * from submissions WHERE year = 2024', dbcon, 'sid')
submissions_raw = pd.read_sql_query('SELECT * from submissions', dbcon, 'sid')
#print(submissions_raw24)

submissions = (submissions_raw
    .join(
        pd.read_sql_query('SELECT * from areas', dbcon, 'aid'), 
        on='aid'
    )
    .assign(Keywords = lambda df: (pd
        .read_sql_query('SELECT * FROM submissionkeywords', dbcon, 'sid')
        .loc[df.index]
        .join(
            pd.read_sql_query('SELECT * FROM keywords', dbcon, 'kid'), 
            on='kid'
        )
        .keyword
        .groupby('sid')
            .apply(list)
    ))
    .assign(**{'# Keywords': lambda df: df.Keywords.apply(len)})
    .assign(**{'FinalDecision': lambda df: df['decision']})
    .replace(staticdata)
    .rename(columns = staticdata['colnames'])
    .drop(columns = ['legacy', 'aid'])
#    .set_index('sid')
#    .set_index('Paper ID')
# note -- I changed the index, since 'Paper ID' was not unique for multiple years.
# By not setting the index to 'Paper ID' the index remains with 'sid'.
# However, 'sid' is used as a unique index in the creation of the database anyways.
)

# replace the old 'Paper ID' with a unique identifier, so that the code from 2021 will work
submissions = submissions.rename(columns = {'Paper ID':'Old Paper ID'})
submissions.reset_index(inplace=True)
submissions['Paper ID'] = submissions['sid']
submissions = submissions.set_index('Paper ID')
#submissions colums: (index), sid (unique id), Paper ID (unique), Old Paper ID, Decision, year, Area, Keywords (as a list), # Keywords

all_years = submissions['year'].unique()

#rates_decision computes the acceptance rates (and total number of papers) per year
#rates_decision: (index), Decision, year, count, Percentage
rates_decision = (submissions
    .value_counts(['Decision', 'year'])
    .reset_index()
    # .rename(columns = {0: 'count'})
)
rates_decision['Percentage'] = rates_decision.groupby(['year'])['count'].transform(lambda x: x/x.sum()*100)
rates_decision = rates_decision.round({'Percentage': 1})
#rates_decision computes the acceptance rates (and total number of papers) per year
#rates_decision: (index), Decision, year, count, Percentage
rates_decision_final = (submissions
    .value_counts(['FinalDecision', 'year'])
    .reset_index()
    # .rename(columns = {0: 'count'})
)
rates_decision_final['Percentage'] = rates_decision_final.groupby(['year'])['count'].transform(lambda x: x/x.sum()*100)
rates_decision_final = rates_decision_final.round({'Percentage': 1})
#submissions
#bids_raw: (index), Reviewer ID, sid (unique paper identifier over mult years), match score, bid of the reviewer, role of the reviewer, Paper ID
bids_raw = (pd
    .read_sql_query('SELECT * from reviewerbids', dbcon)
    .merge(submissions_raw['confsubid'], on='sid')
    .replace(staticdata)
    .rename(columns = staticdata['colnames'])
)
#bids_raw

## Renaming Paper ID to Old Paper ID, setting Paper ID to sid, keeping all 3 for now...
bids_raw = bids_raw.rename(columns = {'Paper ID':'Old Paper ID'})
bids_raw['Paper ID'] = bids_raw['sid']
# bids = Reviewer, sid, Bid (how the reviewer bid on this paper)
#      doesn't include review/sid that were not bid for [.query('Bid != "no bid"')]
bids = (bids_raw
    .query('Bid != "no bid"')
# Paper ID is not unique over multiple years!
#    .drop(columns = ['sid'])
#    [['Reviewer','Paper ID', 'Bid']]
    [['Reviewer','sid', 'Paper ID', 'Bid']]
    .reset_index(drop = True)
)

# matchscores becomes a table to reviewer/sid with the match scores
# many of these will be "NaN" since we now have multiple years together.
# we need to check whether the reviewer IDs remain unique across the years!
matchscores = (bids_raw
# Paper ID is not unique over multiple years!
#    [['Reviewer','Paper ID','match']]
    [['Reviewer','sid','Paper ID','match']]
# Paper ID is not unique over multiple years!
#    .set_index(['Reviewer', 'Paper ID'])
    .set_index(['Reviewer', 'Paper ID'])
    .match
    .unstack(level=1)
)

# assignments = Reviewer, sid, Role (primary, secondary)
#      doesn't include review/sid that were not assigned [.query('Role != ""')]
assignments = (bids_raw
    .query('Role != ""')
# Paper ID is not unique over multiple years!
#    [['Reviewer', 'Paper ID', 'Role']]
    [['Reviewer', 'sid', 'Paper ID', 'Role']]
    .reset_index(drop = True)
)

del dbcon

#### Plot Defaults

acc_template = go.layout.Template()

acc_template.layout = dict(
    font = dict( 
        family='Fira Sans',
        color = 'black',
        size = 13
    ),
    title_font_size = 14,
    plot_bgcolor = 'rgba(255,255,255,0)',
    paper_bgcolor = 'rgba(255,255,255,0)',
    margin = dict(pad=10),
    xaxis = dict(
        title = dict( 
            font = dict( family='Fira Sans Medium', size=13 ),
            standoff = 10
        ),
        gridcolor='lightgray',
        gridwidth=1,
        automargin = True,
        fixedrange = True,
    ),
    yaxis = dict(
        title = dict( 
            font = dict( family='Fira Sans Medium', size=13 ),
            standoff = 10,
        ),
        gridcolor='lightgray',
        gridwidth=1,
        automargin = True,
        fixedrange = True,
    ),
    legend=dict(
        title_font_family="Fira Sans Medium",
    ),
    colorway = px.colors.qualitative.T10,
    hovermode = 'closest',
    hoverlabel=dict(
        bgcolor="white",
        bordercolor='lightgray',
        font_color = 'black',
        font_family = 'Fira Sans'
    ),
)

acc_template.data.bar = [dict(
    textposition = 'inside',
    insidetextanchor='middle',
    textfont_size = 12,
)]

px.defaults.template = acc_template

px.defaults.category_orders = {
    'Decision': list(staticdata['decision'].values()),
    'FinalDecision':  list(staticdata['FinalDecision'].values()),
    'Area': list(staticdata['area'].values()),
    'Short Name': staticdata['keywords']['Short Name'].tolist(),
}

config = dict(
    displayModeBar = False,
    scrollZoom = False,
    responsive = False
)

def aspect(ratio):
    return { 'width': width, 'height': int(ratio*width) }

# useful data sub-products

#k_all columns: (index), Paper ID, Old Paper ID, Decision, year, Area, Keywords (as a list), # Keywords, Keyword, Category, Subcategory, Short Name, Description
k_all = (submissions
    .join(submissions['Keywords']
        .explode()
        .rename('Keyword')
    )
    .reset_index(level = 0)
    .merge(staticdata['keywords'], on='Keyword')
)

# (Old) Paper ID is not unique, however, the 'sid' is (which is the current index)
#k_all.reset_index(inplace=True)
#k_all.rename(columns = {'sid':'Paper ID'},inplace = True)
#k_all = k_all.merge(staticdata['keywords'], on='Keyword')
#k_all

#k_total columns: Category, Subcategory, Short Name, Keyword, Description, #Submissions, year
#  counts the total number of submissions per keyword and year
k_total = staticdata['keywords'].merge(
    k_all.value_counts(['Short Name','year'])
         .rename('# Submissions')
         .reset_index(),
#    on = 'Short Name',
    how = 'right'
#    how = 'outer'
)

#k_cnt: how often was a particular keyword used among all submissions within a year????
#k_cnt columns: (index), Short Name, year, c, Category, Subcategory, Keyword, Description
# not clear how k_cnt and k_total differ!
k_cnt = (k_all
    .value_counts(['Short Name','year'], sort=False)
    .rename('c')
    .to_frame()
    .reset_index()
    .merge(staticdata['keywords'], on='Short Name')
)

Deeper data investigation

This report is generated by members of the ACC for the current year, and prepared for the VSC. Upon review, it will be linked from the IEEE VIS website. The conclusions and discussion points are based on submission and reviewer data from IEEE VIS 2024 (and previous years). The report and analysis performed is focused on the use of keywords, areas, and reviewer matching. Thus, there are likely other aspects of conference organization which are not covered (but could be considered).

The report is broken down into the following sections. After the summary at the beginning, the data and analysis process is described. It shows which data we used, where it is stored, and how it is obtained. These processes can be adapted for future years of this committee.

(NB: Some of the plots shown above are repeated here from the highlights for the sake of completeness.)

Data and Process

We analyzed anonymized data containing information about the full paper submissions to VIS 2024, the reviews of these submissions, and the IPC bidding preferences. We analyzed this data to understand how well the areas and keywords characterize the body of work submitted this year. We also analyzed the IPC bidding information to understand how well the expertise of the IPC members covers the submissions. Below, we show highlights of our findings.

Note that in the the analysis that follows, the submission/paper IDs and reviewer IDs are anonymized through a randomizer, and are not the IDs used in PCS submissions and reviewers.

The data used to perform this analysis is a combination of paper submission data and reviewer bidding data. Both sets were anonymized to minimize the ability to identify IPC members, authors, or reviewers. The analysis of the data in this year uses the anonymized CSV files obtained directly from PCS. You can see the source code used to process and generate the plots in this document by clicking on the “Code” buttons, which will fold out the Python code used. The anonymization script that was used is located in the anonymization-scripts folder (and may be needed to be updated to correspond with changes made in PCS). In order to get ALL the data, it is current run by James at PCS who sends the resultant anonymized files to the ocmmittee where they are stored in the corresponding year folder.

In order to facilitate longitudinal studies of this data, we are also providing a sqlite database with the 2020, 2021, and 2022 data in an attempt to make it easier to incorporate future years. This database (as well as the source code of this document) can be found here

Sanity Checks

We include some sanity checks on the data in order to make sure the data has been processed correctly. In 2024, we should have:

  • 139 papers accepted after the second round
  • 236 papers rejected after the first round
  • 9 papers desk rejected
Code
#rates_decision computes the acceptance rates (and total number of papers) per year
#rates_decision: (index), Decision, year, count, Percentage
rates_decision = (submissions
    .value_counts(['Decision', 'year'])
    .reset_index()
    # .rename(columns = {0: 'count'}) #[jdf] no need to rename, the count is already in the 'count' attribute.
)
rates_decision['Percentage'] = rates_decision.groupby(['year'])['count'].transform(lambda x: x/x.sum()*100)
rates_decision = rates_decision.round({'Percentage': 1})
#rates_decision computes the acceptance rates (and total number of papers) per year
#rates_decision: (index), Decision, year, count, Percentage
rates_decision_final = (submissions
    .value_counts(['FinalDecision', 'year'])
    .reset_index()
    # .rename(columns = {0: 'count'}) #[jdf] no need to rename, the count is already in the 'count' attribute.
)
rates_decision_final['Percentage'] = rates_decision_final.groupby(['year'])['count'].transform(lambda x: x/x.sum()*100)
rates_decision_final = rates_decision_final.round({'Percentage': 1})
#| output: true
rates_decision_final.sort_values(by=['year', 'FinalDecision'], ascending=[False, True], ignore_index=True)

fig = px.bar(rates_decision,
    x = 'count',
    y = 'year',
    barmode = 'stack',
    orientation = 'h',
    color = 'Decision',
    text = 'count',
    custom_data = ['Decision'],
).update_layout(
    yaxis=dict(autorange="reversed"),
    title = 'Submissions',
    xaxis_title = 'Number of Submissions',
    **aspect(0.45)
).update_traces(
    hovertemplate = '%{x} submissions in %{y} have decision %{customdata[0]}<extra></extra>',
).show(config=config)

fig = px.bar(rates_decision,
    x = 'Percentage',
    y = 'year',
    barmode = 'stack',
    orientation = 'h',
    color = 'Decision',
    text = 'Percentage',
    custom_data = ['Decision','count'],
).update_layout(
    yaxis=dict(autorange="reversed"),
    title = 'Submissions',
    xaxis_title = 'Percentage of Submissions',
    **aspect(0.45)
).update_traces(
    hovertemplate = '%{customdata[1]} submissions in %{y} have decision %{customdata[0]}<extra></extra>',
).show(config=config)

Submissions per Area.

We wanted to understand how submissions were distributed by area, including acceptance decisions. Submissions to each area were within reasonable upper and lower limits, and decisions did not appear partial to any individual area.

Code

def group_stat(g):
    return pd.DataFrame({
        '# Submissions': g,
        '% Submissions': round(g/g.sum()*100,1),
        'Total': g.sum()
    })

tmp = (submissions[submissions.year > 2020]
    .value_counts(['Area', 'Decision', 'year'])
    .reset_index()
    .rename(columns = {0: 'count'})
)

fig = px.bar(tmp,
    x = 'count',
    y = 'Area',
    barmode = 'stack',
    orientation = 'h',
    color = 'Decision',
    text = 'count',
    custom_data = ['Decision'],
    facet_row='year',
    category_orders={'year': [2024,2023,2022, 2021]}, #, 2020]},
).update_layout(
    title = 'Submissions by area and year',
    xaxis_title = 'Number of Submissions',
    yaxis=dict(
        autorange="reversed",
        tickfont=dict(size=12),  # Adjust y-label fontsize 
    ),
    **aspect(1.3)
).update_traces(
    hovertemplate = '%{x} submissions in %{y} have decision %{customdata[0]}<extra></extra>',
    texttemplate='%{text}',
    textangle=0  # Force labels to have horizontal orientation
).show(config=config)

fig = px.bar(tmp,
    x = 'count',
    y = 'Area',
    barmode = 'stack',
    orientation = 'h',
    color = 'Decision',
    text = 'count',
    custom_data = ['Decision'],
).update_layout(
    title = 'Submissions by area all years',
    xaxis_title = 'Number of Submissions all years',
    yaxis=dict(
        autorange="reversed",
        tickfont=dict(size=12),  # Adjust y-label fontsize 
    ),
    **aspect(0.5)
).update_traces(
    hovertemplate = '%{x} submissions in %{y} have decision %{customdata[0]}<extra></extra>',
).show(config=config)

Submissions and Keywords used

Keywords with strong variations between 2023 and 2024 are listed here, with their historical differences. Despite the variations, there is no strong trend over the four years. Mostly yearly variations.

Code
k_total['Submission %'] = k_total.groupby(['year'])['# Submissions'].transform(lambda x: x/x.sum()*100)
k_total['Year'] = k_total['year'].astype(str)  # to get categorical colors
k_year = k_total.pivot(index="year", values="Submission %", columns="Short Name").T

pd.set_option('display.max_rows', None)
kw_df = k_year.diff(axis=1).sort_values(by=2024, ascending=False)
kw_df.drop([2020], axis=1, inplace=True)
# maxval = max(-kw_df.min(axis=None), kw_df.max(axis=None))
kw_df[kw_df[2024].abs() > 0.3].style\
    .format('{:.2f}', na_rep="")\
    .text_gradient(cmap="RdYlBu", vmin=-1, vmax=1)
year 2021 2022 2023 2024
Short Name        
MLStatsModel 0.57 -0.24 -0.39 0.78
HumanQual -0.24 0.32 -0.01 0.70
ML 0.45 0.36 -0.73 0.70
LifeBio -0.00 0.18 -0.34 0.65
VisDesign 0.56 -0.06 -0.24 0.59
Text -0.04 -0.38 0.33 0.57
Perception -0.50 0.32 -0.16 0.35
ImageVideo 0.99 -0.33 -0.07 0.31
OtherApp 0.18 -0.38 0.23 -0.31
Time -1.53 0.36 -0.33 -0.33
Application -0.63 0.26 -0.20 -0.37
Algorithm -1.24 0.41 -0.12 -0.45
MultiRes 0.01 -0.47 0.28 -0.45
DataAbstr -0.34 -0.00 0.27 -0.54
Models 0.21 0.04 0.04 -0.57
Code
# do a manual histogram to include non-specified keywords

px.bar(k_total,
    x = 'Short Name',
    y = '# Submissions',
    color = 'Category',
    facet_row='year',
    category_orders={'year': reversed([2020, 2021, 2022, 2023, 2024])},
).update_traces(
    hovertemplate = "'%{x}' specified in %{y} submissions<extra></extra>",
).update_layout(
    xaxis_tickfont_size = 8,
    xaxis_dtick = 1,
    yaxis_dtick = 50,
    hovermode = 'closest',
    title = 'Frequency of keywords across submissions',
    **aspect(0.8)
).show(config=config)

How are keywords distributed across areas?

Code
# do a manual histogram to include non-specified keywords
k_cnt = staticdata['keywords'].merge(
    pd.DataFrame(staticdata['area'].values(), columns = ['Area']), 
    how = 'cross'
).merge(
    k_all
        .value_counts(['Short Name', 'Area'])
        .rename('# Submissions')
        .reset_index(),
    how = 'outer'
).fillna(1e-10) # needed for sorting, Plotly bug?

px.bar(k_cnt,
    x = 'Short Name',
    y = '# Submissions',
    color = 'Area',
    custom_data = ['Area']
).update_traces(
    hovertemplate = 'Keyword "%{x}" specified by %{y} submissions from area "%{customdata}"<extra></extra>'
).update_layout(
    barmode = 'stack',
    xaxis_dtick = 1,
    xaxis_tickfont_size = 8,
    xaxis_fixedrange = True,
    yaxis_fixedrange = True,
    xaxis_categoryorder = 'total descending',
    title = 'Frequency of keywords across submissions, by area',
    **aspect(0.5)
).show(config=config)

How many submissions specified a given number of keywords?

Code
tmp = (submissions
    .value_counts(['# Keywords', 'Area'])
    .rename('# Submissions')
    .reset_index()
)

px.bar(tmp,
    x = '# Keywords', 
    y = '# Submissions',
    barmode = 'stack',
    color = 'Area',
    custom_data=['Area'],
).update_traces(
    hovertemplate = '%{y} submissions specified %{x} keywords in area "%{customdata}"<extra></extra>',
).update_layout(
    xaxis_dtick = 1,
    title = 'Keyword count per submission',
    **aspect(0.5)
).show(config=config)

How often are pairs of keywords specified together?

Code
k_pairs = (k_all
    .groupby('Paper ID', )
    .apply(lambda g: pd.DataFrame(itertools.combinations(g['Short Name'].values, 2)),
          include_groups=False)
    .join(submissions['Decision'])
)

tmp = k_pairs.groupby([0,1]).size().nlargest(40)
tmp = (
    k_pairs
    .set_index([0,1])
    .loc[tmp.index]
    .assign(**{'Keyword Pair': lambda df: [' + '.join(v) for v in df.index.values]})
    .value_counts(['Keyword Pair', 'Decision'], sort=False)
    .rename('# Submissions')
    .reset_index()
)

cooc = (k_pairs
    .groupby([0,1])
    .size()
    .unstack()
    .reindex(
        index = staticdata['keywords']['Short Name'], 
        columns = staticdata['keywords']['Short Name']
    )
    .fillna(0)
)

cooc = (cooc + cooc.T)
cooc['Total'] = cooc.max(axis=0)
cooc.sort_values(by='Total', ascending=False,inplace=True)
cooc = cooc.loc[:,list(cooc.index)]

np.fill_diagonal(cooc.values, None)

px.imshow(cooc, 
    color_continuous_scale='portland',
).update_traces(
    connectgaps = False,
    hoverongaps = False,
    hovertemplate = "Keywords '%{x}' and '%{y}' are jointly specified in %{z} submissions<extra></extra>",
    colorbar_title = '# Submissions',
).update_layout(    
    xaxis_dtick = 1,
    xaxis_tickfont_size = 7,
    yaxis_dtick = 1,
    yaxis_tickfont_size = 7,
    hovermode = 'closest',
    xaxis_showgrid = False,
    yaxis_showgrid = False,
    title = 'Co-occurrence of keywords',
    **aspect(.7)
).show(config=config)

Matching for reviewers

What is the distribution of match scores by keyword?

Code
tmp = (matchscores.T
    .stack()
    .rename('Score')
    .loc[lambda x: x > -1.0]
    .reset_index()
    .merge(k_all.loc[:,['Paper ID', 'Short Name', 'Category']], on='Paper ID')
    #.merge(k_all.loc[:,['sid', 'Short Name', 'Category']], on='sid')
    .reset_index()
)

px.box(tmp,
    x = 'Short Name',
    y = 'Score',
    color = 'Category',
).update_layout(
    xaxis_dtick = 1,
    xaxis_tickfont_size = 8,
    title = 'Distribution of match scores per keyword',
    **aspect(0.4)
).update_traces(
    width = .5,
    line_width = 1,
).show(config=config)

What is the number of “high” match scores, per submission?

Code
tmp = (matchscores.T
    .where(lambda x: x > -1.0, None)
)

threshold = [0.5, 0.7, 0.9]

tmp = pd.concat([ 
    (tmp >= q).agg('sum', axis=1).rename('≥ %.1f' % q) 
    for q in threshold 
], axis=1)

px.bar(tmp,
    barmode = 'overlay',
    opacity = 1,
).update_layout(
    legend_title = 'Match Score',
    bargap = 0.1,
    xaxis_type = 'category',
    xaxis_tickfont_size = 8,
    yaxis_title = '# Reviewers',
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Submission %{x} has %{y} matches %{fullData.name}<extra></extra>'
).show(config=config)

Which are the 10 submissions with the lowest number of “good” (≥ 0.5) match scores?

Code
ind = tmp['≥ 0.5'].nsmallest(10).index

(tmp
    .merge(bids
        .query('`Paper ID` in @ind and Bid in ["willing", "want"]')
        .value_counts(['Paper ID'])
        .rename("Pos. Bids"),
        on='Paper ID')
    .merge(submissions, on='Paper ID')
    .sort_values('≥ 0.5')
    .assign(Keywords = lambda df: df.Keywords
        .apply(lambda x: ', '.join(
            [staticdata['keywords'].set_index('Keyword').loc[k, 'Short Name'] for k in x]
    )))
)
≥ 0.5 ≥ 0.7 ≥ 0.9 Pos. Bids sid Old Paper ID Decision year Area Keywords # Keywords FinalDecision
Paper ID
873 0 0 0 14 873 287 Reject 2021 Theoretical & Empirical OtherContrib, OtherTopic 2 Reject
1139 0 0 0 16 1139 111 Accept 2022 Theoretical & Empirical NAData, OtherContrib 2 Accept
2525 32 4 0 4 2525 498 Reject Strong 2024 Representations & Interaction Color, Art, ImageProcessing 3 Reject
2289 38 27 11 12 2289 262 Accept 2024 Data Transformations Scalar, CompTop, Isosurfaces 3 Accept
2360 41 21 14 10 2360 333 Reject 2024 Systems & Rendering Isosurfaces, Flow 2 Reject
2357 43 25 13 14 2357 330 Reject 2024 Representations & Interaction Vector_Tensor, Features, CompTop, Flow 4 Reject
2045 44 18 0 2 2045 18 Reject Strong 2024 Systems & Rendering Systems, Software, BigData, Isosurfaces, Volumes 5 Reject
2292 44 15 0 4 2292 265 Reject 2024 Applications ImageVideo, Workflow, Displays 3 Reject
2464 44 28 12 14 2464 437 Accept 2024 Data Transformations Vector_Tensor, Algorithm, Features, CompTop, Flow 5 Accept
2237 45 26 8 7 2237 210 Accept 2024 Data Transformations Scalar, BigData, CompTop 3 Accept

What is the number of “high” match scores, per keyword?

Code
tmp = (matchscores.T
    .where(lambda x: x > -1.0, None)
)

tmp = (k_all
    .set_index('Paper ID')['Short Name']
    .to_frame()
    .merge(tmp, left_index=True, right_index=True, how='inner')
    .set_index('Short Name')
)

tmp = (pd
    .concat([ 
        (tmp >= q).agg('sum', axis=1).rename('≥ %.1f' % q) 
        for q in threshold 
    ], axis=1)
    .groupby('Short Name')
    .mean()
)

px.bar(tmp,
    barmode = 'overlay',
    opacity = 1,
).update_layout(
    legend_title = 'Match Score',
    bargap = 0.1,
    xaxis_dtick = 1,
    xaxis_type = 'category',
    xaxis_tickfont_size = 8,
    yaxis_title = '# Matches',
    title = 'Match scores by keyword',
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Keyword %{x} has %{y:.1f} matches %{fullData.name}<extra></extra>'
).show(config=config)



px.bar(tmp.sort_values('≥ 0.9', ascending=False),
    barmode = 'overlay',
    opacity = 1,
).update_layout(
    legend_title = 'Match Score',
    bargap = 0.1,
    xaxis_dtick = 1,
    xaxis_type = 'category',
    xaxis_tickfont_size = 8,
    yaxis_title = '# Matches',
    title = 'Match scores by keyword',
    xaxis_categoryorder = "trace",
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Keyword %{x} has %{y:.1f} matches %{fullData.name}<extra></extra>'   
).show(config=config)


px.bar(tmp.sort_values('≥ 0.7', ascending=False),
    barmode = 'overlay',
    opacity = 1,
).update_layout(
    legend_title = 'Match Score',
    bargap = 0.1,
    xaxis_dtick = 1,
    xaxis_type = 'category',
    xaxis_tickfont_size = 8,
    yaxis_title = '# Matches',
    title = 'Match scores by keyword',
    xaxis_categoryorder = "trace",
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Keyword %{x} has %{y:.1f} matches %{fullData.name}<extra></extra>'   
).show(config=config)


px.bar(tmp.sort_values('≥ 0.5', ascending=False),
    barmode = 'overlay',
    opacity = 1,
).update_layout(
    legend_title = 'Match Score',
    bargap = 0.1,
    xaxis_dtick = 1,
    xaxis_type = 'category',
    xaxis_tickfont_size = 8,
    yaxis_title = '# Matches',
    title = 'Match scores by keyword',
    xaxis_categoryorder = "trace",
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Keyword %{x} has %{y:.1f} matches %{fullData.name}<extra></extra>'   
).show(config=config)

Bidding

How many bids did individual PC members make?

Code
tmp = (bids
    .value_counts(['Reviewer', 'Bid'], sort=False)
    .rename('# Bids')
    .reset_index()
)

px.bar(tmp,
    x = 'Reviewer',
    y = '# Bids',
    color = 'Bid'
).update_layout(
    xaxis_type = 'category',
    xaxis_categoryorder = 'total descending',
    xaxis_showticklabels = False,
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Reviewier %{x} made %{y} "%{fullData.name}" bids.<extra></extra>'
).show(config=config)

How many (positive) bids did each submission receive?

Code
tmp = (bids
    .value_counts(['Paper ID', 'Bid'], sort=False)
    #.value_counts(['sid', 'Bid'], sort=False)
    .rename('# Bids')
    .reset_index()
    .loc[lambda x: x.Bid.isin(['want', 'willing'])]
)

px.bar(tmp,
    x = 'Paper ID',
    #x = 'sid',
    y = '# Bids',
    color = 'Bid'
).update_layout(
    xaxis_type = 'category',
    xaxis_categoryorder = 'total descending',
    xaxis_showticklabels = False,
    title = 'Positive Bids per Paper',
    **aspect(0.4),
).update_traces(
    hovertemplate = 'Paper %{x} received %{y} "%{fullData.name}" bids.<extra></extra>',
).show(config=config)
Code
popular = 15

tmp = (bids
    .query('Bid in ["want", "willing"]')
    .value_counts(['Paper ID', 'Bid'], sort=False)
   # .value_counts(['sid', 'Bid'], sort=False)
    .unstack()
    .fillna(0)
    .groupby(['want', 'willing'])
    .apply(lambda g: pd.Series({'ids': g.index.values, 'count': g.index.size}),
          include_groups=False)
    .reset_index()
    .assign(popular = 
            lambda df: np.where( df['willing']+df['want']>=popular, "≥ %d" % popular, "< %d" % popular)
))

px.scatter(tmp,
    x = 'willing',
    y = 'want',
    size = 'count',
    color = 'popular',
    custom_data = ['count', 'ids'],
).update_layout(
    legend_title = 'Total Pos. Bids',
    title = 'Distribution of Positive Bids',
    **aspect(0.4)
).update_traces(
    hovertemplate = '%{customdata[0]} papers received %{x} "willing" and %{y} "want" bids',
).show(config=config)

Does the presence of specific keywords correlate with bidding?

We run a reviewer-independent ridge regression model where the independent variable is the overall reviewer interest, and the dependent variable is the (weighted) presence of a keyword. We measure interest by giving each “willing” or “want” bid a score of 1:

Code
tmp_3 = staticdata['keywords'].copy()
tmp_3['ix'] = list(range(len(tmp_3)))
tmp_3 = tmp_3[['Short Name', 'ix']]
tmp_1 = k_all[['Paper ID', 'Short Name']]
#tmp_1 = k_all[['sid', 'Short Name']]
tmp_2 = bids[(bids['Bid'] == 'willing') | (bids['Bid'] == 'want')]
df = tmp_1.merge(tmp_3, on="Short Name").merge(tmp_2, on="Paper ID")
#df = tmp_1.merge(tmp_3, on="Short Name").merge(tmp_2, on="sid")
df['weight'] = 2
df.loc[df['Bid'] == 'willing', 'weight'] = 1

total_weight = df[['Paper ID', 'ix', 'weight']].groupby(['Paper ID', 'ix']).sum().reset_index()
keyword_count = tmp_1.groupby(['Paper ID']).count().reset_index()
#total_weight = df[['sid', 'ix', 'weight']].groupby(['sid', 'ix']).sum().reset_index()
#keyword_count = tmp_1.groupby(['sid']).count().reset_index()
keyword_count['Keyword Weight'] = 1.0/keyword_count['Short Name']
total_weight = total_weight.merge(keyword_count[['Paper ID', 'Keyword Weight']], on="Paper ID")
nrows = max(total_weight['Paper ID']) + 1
#total_weight = total_weight.merge(keyword_count[['sid', 'Keyword Weight']], on="sid")
#nrows = max(total_weight['sid']) + 1

ncols = max(total_weight['ix']) + 1
design_matrix = np.zeros((nrows, ncols))
design_matrix.shape
rhs = np.zeros(nrows)

# this is embarrassing, there must be a fancy pandas way of doing it.
# someone else can figure it out.
for i, row in total_weight.iterrows():
    design_matrix[int(row['Paper ID']), int(row['ix'])] = row['Keyword Weight']
    #design_matrix[int(row['sid']), int(row['ix'])] = row['Keyword Weight']
    rhs[int(row['Paper ID'])] = row['weight']
    #rhs[int(row['sid'])] = row['weight']
import scipy.linalg
from sklearn.linear_model import Ridge
# Ideally, we find the best regularizer by splitting into training/validation,
# but on inspection the order doesn't seem to change too much 
lr = Ridge(1).fit(design_matrix, rhs)
lr.coef_
tmp_3['Importance'] = lr.coef_
tmp_3 = tmp_3.sort_values(by=['Importance']).merge(staticdata['keywords'], on='Short Name', )

px.scatter(tmp_3, 
    x="Short Name", 
    y="Importance", 
    color='Category',
    custom_data = ['Keyword'],
).update_layout(
    title = 'Keyword Importance for Bidding',
    xaxis_dtick = 1,
    xaxis_categoryorder = 'trace',
    xaxis_tickfont_size = 8,
    **aspect(0.4)
).update_traces(
    hovertemplate = 'Importance of "%{customdata[0]}": %{y}<extra></extra>'
).show(config=config)

Assignment

How many papers were PC members assigned?

Code
tmp = assignments.value_counts(['Reviewer']).rename('# Assignments').reset_index()


px.histogram(tmp,
    x = '# Assignments',
).update_traces(
    hovertemplate = '%{y} reviewers were assigned %{x} submissions',
).update_layout(
    bargap = .1,
    yaxis_title = '# PC members',
    title = 'Distribution of assignments',
    
    **aspect(0.4)
).show(config=config)

### TODO: Split by year 

tmp2=assignments.merge(submissions['year'], on='Paper ID').value_counts(['Reviewer', 'year']).rename('# Assignments').reset_index()

px.histogram(tmp2,
    x = '# Assignments',
    facet_row="year",
    category_orders={'year': [2023, 2022, 2021]}
).update_traces(
    hovertemplate = '%{y} reviewers were assigned %{x} submissions'
).update_layout(
    bargap = .1,
    yaxis1_title = '# PC members',
    yaxis2_title = '# PC members',
    yaxis3_title = '# PC members',
    title = 'Distribution of assignments, per year',
    **aspect(0.5)
).show(config=config)

How many areas did reviewers review in?

Code
tmp = (assignments
    .merge(submissions, on='Paper ID')
    #.merge(submissions, on='sid')
    .groupby('Reviewer')
    .apply(lambda x: len(x['Area'].unique()),
          include_groups=False)
    .reset_index())

px.histogram(tmp,
    x = 0,
).update_traces(
    hovertemplate = '%{y} PC members were assigned submissions from %{x} area(s)',
).update_layout(
    bargap = .1,
    xaxis_title = '# Areas',
    yaxis_title = '# PC members',
    **aspect(0.4),
).show(config=config)

### TODO: Split by year

tmp = (assignments
    .merge(submissions, on='Paper ID')
    .groupby(['Reviewer', 'year'])
    .apply(lambda x: len(x['Area'].unique()),
          include_groups=False)
    .reset_index())

px.histogram(tmp,
    x = 0,
    facet_row='year',
    category_orders={'year': [2023, 2022, 2021]}
).update_traces(
    hovertemplate = '%{y} PC members were assigned submissions from %{x} area(s)',
).update_layout(
    bargap = .1,
    xaxis_title = '# Areas',
    yaxis1_title = '# PC members',
    yaxis2_title = '# PC members',
    yaxis3_title = '# PC members',
    **aspect(0.5),
).show(config=config)

How do match scores correlate with bids?

Code
### TODO: Split by year

tmp = bids.assign(
    Score = bids.apply(lambda x: (matchscores.loc[x['Reviewer'], x['Paper ID']]), axis=1),
    Area  = bids.apply(lambda x: (submissions.loc[x['Paper ID'], 'Area']), axis=1), 
    year  = bids.apply(lambda x: (submissions.loc[x['Paper ID'], 'year']), axis=1), 
).query('Score > -1.0')

px.box(tmp,
    x = 'Bid',
    y = 'Score',
    color = 'Bid',
    facet_row='year'
).update_layout(
    showlegend = False,
    xaxis_categoryorder = 'array',
    xaxis_categoryarray = ['want', 'willing', 'reluctant', 'conflict'],
    **aspect(0.8)
).update_traces(
    line_width = 2,
    boxmean = True
).show(config=config)
Code
px.violin(tmp,
    x = 'Bid',
    y = 'Score',
    color = 'Area',
    box = True,
).update_layout(
    # showlegend = False,
    title = 'Match scores by bid by area',
    xaxis_categoryorder = 'array',
    xaxis_categoryarray = ['want', 'willing', 'reluctant', 'conflict'],
    violingap=0.2, 
    violingroupgap=0.1,
    **aspect(0.4)
).update_traces(
    box_line_color = 'black',
    box_line_width = 1,
    line_width = 0,
    meanline_visible=True,
    marker_size = 4,
    # boxpoints = 'outliers',
).show(config=config)

### TODO: Split by year

px.violin(tmp,
    x = 'Bid',
    y = 'Score',
    color = 'Area',
    box = True,
    facet_row='year'
).update_layout(
    # showlegend = False,
    title = 'Match scores by bid by area, by year',
    xaxis_categoryorder = 'array',
    xaxis_categoryarray = ['want', 'willing', 'reluctant', 'conflict'],
    violingap=0.2, 
    violingroupgap=0.1,
    **aspect(0.8)
).update_traces(
    box_line_color = 'black',
    box_line_width = 1,
    line_width = 0,
    meanline_visible=True,
    marker_size = 4,
    # boxpoints = 'outliers',
).show(config=config)

How often were reviewers assigned submissions that they bid on?

Code
tmp = (
    assignments
        .merge(bids, on=['Reviewer', 'Paper ID'], how='left')
#        .merge(bids, on=['Reviewer', 'sid'], how='left')
        .value_counts(['Role', 'Bid'])
        .rename('Reviewers')
        .reset_index()
)

fig = px.bar(tmp,
    y = 'Reviewers',
    x = 'Role',
    color = 'Bid',
    custom_data = ['Bid']
).update_traces(
    hovertemplate = '%{y} PC members assigned as %{x} bid %{customdata}<extra></extra>',
).update_layout(
    title = "Assignment by bidding",
    **aspect(0.4),
).show(config=config)

### TODO: Split by year
tmp2 = (
    assignments
        .merge(bids, on=['Reviewer', 'Paper ID'], how='left')
        .merge(submissions, on='Paper ID')
        .value_counts(['Role', 'Bid', 'year'])
        .rename('Reviewers')
        .reset_index()
)

fig = px.bar(tmp2,
    y = 'Reviewers',
    x = 'Role',
    color = 'Bid',
    custom_data = ['Bid'],
    facet_row='year'  
).update_traces(
    hovertemplate = '%{y} PC members assigned as %{x} bid %{customdata}<extra></extra>',
).update_layout(
    title = "Assignment by bidding, per year",
    **aspect(0.5),
).show(config=config)